The lab · for fun

One prompt. Many models.

Side projects that got out of hand. Long cinematic prompts run through frontier and local models, every run rendered live in your browser and comparable side by side.

3 prompts26 runs21 models
LLM prompt testsAll prompt tests →
12 runs

Bioluminescent Abyssal Temple

One prompt, twelve model runs: an interactive underwater voxel temple scene with bioluminescent life, god rays, and cinematic camera work.

Qwen 3.8 27B MLX 8bit FP16Qwen 3.8 27B MLX 4bit FP16GLM 5.2 MaxSWE-1.7 MaxGrok 4.6 xHighDeepSeek Flash 0731 Max+6
10 runs

Totality Engine

One prompt, ten model runs: a 32-second looping cinematic — a drowned gothic cathedral at solar-eclipse totality, a monumental brass astronomical clock, directed camera beats, and custom GLSL water, corona, lensing, and thick glass.

GPT-6-Astra xHighFable 5.1 xHighOpus 5.5Ox Alpha xHighDevin FusionKimi K3 Max+4
4 runs

Fresnel Engine

One prompt, four model runs: a 32-second looping cinematic — a four-meter first-order Fresnel lens turning in a storm-lashed lighthouse lantern room, directed camera beats, and custom GLSL honey glass, volume beam, rain, and caustics.

Secret Model MaxKimi K3 MaxGPT-6-Astra xHighOpus 5.5
Interactive

AI tierlist

Drag and rank AI tools and models into tiers.

In progress · not ready yet

LLM Visual Bench

A benchmark for visual output from language models.

Write-up

Ornith 9B failure analysis

Why the Ornith 1.5 9B 8-bit run stays stuck on its loading screen.

Tool

Compare runs

Two model runs side by side, rendered live.